In this study of molecular generation using masked diffusion models (MDMs), we diagnose the performance degradation of existing MDMs as a "state collision" problem and propose Masked Element-wise Learnable Diffusion (MELD) to address this issue by adjusting element-specific decay trajectories. MELD uses a parameterized noise scheduling network that assigns different decay rates to individual graph elements, such as atoms and bonds. Across various molecular benchmarks, MELD improves overall generation quality, increases the chemical validity of conventional MDMs from 15% to 93% on the ZINC250K dataset, and achieves state-of-the-art performance in conditional generation tasks.
